What to Do When Your AI Scribe Mishears a Medication

The error every physician fears

You're reviewing the AI-generated note from your last patient. The plan section says "increased to lisinopril 40 mg daily." But you actually said "we'll increase the lisinopril dose, let me think about whether 20 or 40 is right." The AI heard a conditional discussion and wrote it as a confirmed order.

This is the kind of error that, if signed and acted on, becomes a real patient safety event. The medication is real. The patient is real. The dose is wrong.

This article covers how to catch these errors during your normal review, how to prevent them in the first place, and what to do if one slips through and you discover it after the note is signed.

Where medication errors happen most

Across reported and observed patterns in 2026, medication misheard or miscaptured by AI scribes cluster in four predictable categories:

Sound-alike drug names. Celebrex versus Celexa. Hydroxyzine versus hydralazine. Klonopin versus clonidine. Modern medical speech models have largely solved the easy cases, but specific pairs still trip even the best systems. Roughly 30 to 40 percent of medication errors in AI-generated notes are sound-alike substitutions.

Dose magnitude. "Fifteen" heard as "fifty." "Five milligrams" heard as "twenty-five milligrams" when the audio is unclear. Off-by-one-decimal errors are rare but the highest consequence. Roughly 25 percent of medication errors.

Frequency. "Twice daily" versus "three times daily" versus "as needed." Less consequential than dose errors but more common, especially when the provider says something like "let's try twice a day and see how you tolerate it" and the AI captures only "twice a day" without the conditional. Roughly 20 percent.

Discussion versus order. The kind of error in the opening scenario. The provider was thinking out loud or discussing options with the patient, and the AI couldn't reliably distinguish "I'm considering X" from "I'm prescribing X." Roughly 15 percent.

The remaining 5 to 10 percent are edge cases: route of administration ("PO" versus "IM"), brand versus generic, drug strength versus pill count, and so on.

A review pattern that catches them

Reading every word of every note in equal depth is not realistic at clinic volume. The pattern that works is differential review: skim the parts that rarely contain medication errors, read the parts that often do.

Always read carefully:

The medications section (current meds, new meds, changes)
The plan section, specifically any line that contains a drug name, a dose, a frequency, or a route
Anything ending in "mg," "mcg," "units," or "tab"

Skim:

Chief complaint
History of present illness paragraphs that don't mention medications
Past medical history (assuming it's pulled from EHR, not from the conversation)
Social history
Physical exam findings

Cross-check:

The new medication list in the note against the medication list in your EHR. Any difference is either a real change you ordered (intentional) or an error (catch it).
Frequencies and doses against your verbal instruction to the patient if you can remember.

The differential pattern takes 30 to 60 seconds for a typical note. Reading everything in equal depth takes 2 to 4 minutes. The catch rate is comparable because the medication-prone sections get more attention either way.

What to do when you catch one during review

You're reading the note, you see the AI wrote "lisinopril 40 mg" when you intended to write "lisinopril 20 mg pending labs." Three steps:

Correct the note before signing. This is the easy case. Edit the medication, dose, and any conditional language the AI dropped. Sign.
Note the pattern for later. Most AI scribes let you flag the specific error so the system learns. Even if it doesn't have a built-in mechanism, jot down what you saw. Patterns of error tell you whether to keep using the same review approach or whether something has changed about the system.
If this happened on a high-stakes medication (warfarin, insulin, opioid, controlled substance, anything narrow-therapeutic-index), spend an extra 10 seconds verifying the rest of the note. When the AI gets one medication wrong, the chance of a second error in the same note goes up. The differential review pattern catches most of these, but the high-stakes-prescription cases warrant a second pass.

What to do if one slips through

Less pleasant scenario. You sign the note. The patient picks up the prescription. A week later you discover the AI captured the dose wrong and the prescription went out at the wrong dose.

The playbook:

Within hours of discovery:

Contact the pharmacy. If the prescription hasn't been picked up, intercept it. If it has, the pharmacist needs to know.
Contact the patient. Don't email. Call. Explain clearly: "When I reviewed your chart, I noticed the dose on your prescription was incorrect. Here's the correct dose. Please stop taking the wrong one and pick up the corrected prescription." Apologize without blaming the technology. The patient's trust matters more than the explanation.
Document the correction. In the original note, add an addendum explaining the error and the correction. Most EHRs handle this with an "Addendum" or "Late Entry" feature. Use it; don't edit the original note silently.
If the patient took the wrong dose, assess clinical harm. Most narrow errors don't cause harm. Some do. Treat any actual harm as a patient safety event.

Within days:

Report internally per your practice's incident reporting workflow. Even if no harm occurred, near-misses are tracked.
If your jurisdiction requires reporting to a licensing body for medication errors causing harm, report. The threshold varies. In Canada, provincial Colleges of Physicians and Surgeons have specific reporting frameworks.
Notify your malpractice carrier if there was any harm or any patient complaint. They want to know early, not late.
Notify your AI scribe vendor. Send the specific recording and the specific note. Vendors that take patient safety seriously will investigate the model behavior; vendors that don't will deflect. Both responses tell you something.

Within weeks:

Review whether the error reveals a systematic issue or a one-off. If it's systematic (the AI consistently mishears a specific drug name in your accent or context), change your review pattern to compensate, or escalate with the vendor.
If your malpractice carrier asked for a remediation plan, document one. The plan can be as simple as "all medication-related lines in AI-generated notes will be verified against EHR before signing," with a sign-off process.

How to reduce the chance in the first place

Several configuration choices reduce medication error rates substantially:

Have the AI flag confidence on medication mentions. Some platforms (Transcribe Health does this; some competitors do; many don't) annotate medication mentions with a confidence score. Anything below a high-confidence threshold gets visually marked for explicit review. This adds maybe 5 seconds per low-confidence medication and catches the majority of the residual error.

Have the AI cross-check medications against the patient's existing med list. A drug name that doesn't match anything currently prescribed and isn't on a common-prescription list for the patient's conditions should be flagged. Not all platforms do this; ask before signing up.

Use a microphone with good signal-to-noise. Most medication misheard cases trace back to audio quality. A dedicated lapel mic or a wired desktop mic outperforms a laptop's built-in mic by enough margin to meaningfully change accuracy. The hardware cost is trivial compared to the consequence.

Verbalize doses clearly. Speaking to the patient, you might naturally say "let's increase the lisinopril to forty." Saying "let's increase the lisinopril to four-zero milligrams daily" feels stilted but reduces the dose misheard rate noticeably. Some providers find a middle ground: stating the dose twice, once conversationally and once explicitly: "increase the lisinopril to forty, that's four-zero milligrams daily."

Be explicit about conditionals. Instead of "let's try forty, see how it goes," say "we'll increase to forty milligrams; I'll order it pending the labs." Phrases like "pending" and "if X, then Y" give the AI signal that this is a conditional, not a confirmed order.

What the data says about AI-scribe medication accuracy in 2026

Realistic medication-capture accuracy across the major ambient platforms runs in the 92 to 97 percent range for common medications, and 85 to 93 percent for rare or specialty-specific medications. See clinical NLP accuracy benchmarks for the full breakdown.

Those numbers sound high until you do the math: a practice with 100 medication mentions per day operates at a baseline of 3 to 8 captured-incorrectly medications per day in raw AI output, before physician review. The physician review process is what catches them and brings the residual error rate down to roughly 1 in 1,000 to 1 in 5,000 signed notes containing a medication error.

Compared to human-typed notes (which have their own medication error rate, estimated at 1 in 200 to 1 in 1,000 signed notes), AI-with-review is comparable or modestly better. The risk profile is different (errors are more systematic, less random), and the mitigation pattern is different (review carefully, not retype carefully), but the bottom-line patient safety profile is broadly similar.

The honest framing

AI scribes do introduce a new class of error. Patient safety isn't worse than human-typed notes on average, but the failure modes are different and the mitigation depends entirely on the physician review step. Skip the review and you've introduced risk without any of the controls. Do the review well and you're at least as safe as before.

If you're starting with an AI scribe, build the review step into your habit from day one, not month three. The pattern that works is consistent, not perfect: skim what's safe, read what's risky, catch what matters.

For the broader question of accuracy across all dimensions of clinical NLP (not just medications), see clinical NLP accuracy benchmarks for 2026. For how AI scribes handle the review-and-correction workflow specifically, review and approve AI clinical notes covers the mechanics.